10 AI Features in Modern Smartphones, and What Runs Where

Phones now ship with features that translate a phone call as you speak, summarise a recording, remove a stranger from a photo and caption anything with sound. They are described with similar marketing words by Apple, Google and Samsung, but the useful question is practical: where does the work happen, and what does the feature see while it runs?

This guide describes ten features across the three platforms, using each vendor's own wording and pages. It does not rank them, test them or say which phones have them. It exists so that you know what to look for before you switch one on.

What do the vendors say about where AI runs?

Apple. Apple's Apple Intelligence page says the system uses 'powerful on-device processing' and 'can draw on larger server-based models, without making your data accessible to Apple or anyone else.' Its Private Cloud Compute design post, dated June 10, 2024, says Apple prefers on-device computation where possible, sends demanding requests to its own servers, and states that 'no user data is retained in any form after the response is returned.' That post describes design goals and commitments, and the page itself includes no independent verification.

Google. Google's Pixel 10 feature post, dated August 20, 2025, says Tensor G5 'is the first chip to run our newest Gemini Nano model' and describes several features as running on-device. Individual help pages then say more feature by feature. Call Notes is described as using an on-device Gemini model, while Recorder summaries run in the cloud.

Samsung. Samsung's Galaxy AI page says: 'You can decide where your Galaxy AI data gets processed — either locally on device or out in the cloud.' It points to privacy settings for that choice. The feature descriptions on the page generally do not say which setting applies to which feature.

Ten features at a glance

1. Live Translation on iPhone: what does Apple say?

Apple describes Live Translation as translating text in Messages and speech in Phone and FaceTime calls. In its support article, published September 14, 2026, Apple states: 'Live Translation runs entirely on device to keep your messages private.' In calls, both people are alerted that translation is on, and FaceTime shows translated live captions. Languages are downloaded first, and Apple notes that features 'might not be available in all regions or languages.'

Limit to watch: Apple's own page warns that generative outputs 'may vary, so check important information for accuracy.' Do not rely on a translated call for medical, legal or financial details without a second check.

2. Live Translate for calls on Pixel: what does Google say?

Google's help page says the call feature is off by default and is started from Call Assist during a call. It adds that 'Both you and the other party will each hear a brief announcement in your own language,' that 'No conversation audio or transcription is stored on the device,' and that 'Conversations aren't sent to Google servers or anywhere else, or retrievable after the call.' The Pixel 10 post says on-device AI translates the call in real time. Google calls it Voice Translate on Pixel 10 and Live Translate on later models.

Limit to watch: translation covers a fixed set of languages that Google lists, and it can still mistranslate. The page does not state outright that all processing is on the device, though the offline claim points that way.

3. Interpreter on Galaxy: what does Samsung say?

Samsung describes Interpreter as translating speech in real time, with audio delivered to Galaxy Buds and text shown on the phone. The page does not say whether Interpreter runs locally or in the cloud; Samsung's general statement is that users choose where Galaxy AI data is processed.

Limit to watch: check your privacy setting before using it for a sensitive conversation, because the processing location is a setting, not a fixed fact about the feature.

4. Call Notes on Pixel: what happens to a recorded call?

Google says Call Notes transcribes calls and creates summaries with key points, dates, times, places and next steps, and keeps a transcript and recording you can open from your call log. According to the help page, 'The call is transcribed and then summarized by an on-device Gemini model,' and 'Your call contents are stored on your device and aren't shared with Google.' It also states that 'participants are notified' when you start it.

Limit to watch: recording laws on consent differ by place. The notification helps, but you are still responsible for knowing the rules where you and the other person are. Google lists language and region limits, so confirm on your device.

5. Recorder on Pixel: which step uses the cloud?

Google's Recorder help page says summaries are created from a transcript and that the summary feature 'runs in a secure, isolated environment in the cloud and keeps all your info protected and private to you.' It adds that 'Certain languages may need internet access for summarization,' and that re-transcribing 'may process audio files on Google servers.' The page we read does not say where real-time transcription runs.

Limit to watch: this is a good example of one app mixing on-device and cloud steps. If a recording is sensitive, skip the summary step, or read the page for the step you plan to use.

6. Transcript Assist on Galaxy: what does it do?

Samsung says Transcript Assist converts recordings to text and can then summarise or translate them. As with Interpreter, the page does not state where each step runs, and Samsung points users to its privacy settings.

Limit to watch: a summary is a compression. It can leave out the one detail you needed. Keep the original recording or transcript until you have checked anything that matters.

7. Clean Up and Photo Assist: removing things from photos

Apple describes Clean Up as a tool that can 'remove even larger objects' from a photo. Samsung describes Photo Assist as letting you 'move, resize, or delete objects' and having AI fill the gaps, and also add new elements with a prompt or reference image. Neither page we read says where the editing is processed.

Limit to watch: a photo with an object removed or added is an edited image. If you share it as a record of what happened, such as for insurance, a dispute or news, say that it was edited. Google's Circle to Search page lists a check for whether photos are original or modified, but notes that it may not work in some apps.

8. Circle to Search: what does it see?

Google's Android help page says you touch and hold the home button or navigation handle to open a search bar, then circle, highlight or tap what you want to look up. It describes whole-screen translation, a photo-authenticity check and AI Overviews. The page says some features are limited to specific models and regions, and that translation 'works for most content types like web pages, messages and social media, but not videos and GIFs.'

Limit to watch: the page has no dedicated privacy statement for the feature. It says history and personalisation stay controlled by Web & App Activity, and its troubleshooting steps mention a 'Use screen and app data' setting. The feature works on what is on your screen, so be aware of that when something sensitive is visible.

9. Live Caption: an accessibility feature that stays on the phone

Google says Live Caption captions speech on your device, including videos, podcasts, phone calls, video calls and audio messages. Its accessibility help page states: 'All audio and captions are processed on the device and are never stored or sent to Google,' and that it does not need mobile data once the default language pack is downloaded. It also says it 'isn't intended for calls with more than one other person' and 'doesn't work on music.'

Limit to watch: captions are machine-generated and can be wrong, especially with accents, background noise or technical terms. Apple's Apple Intelligence page also lists accessibility enhancements, such as VoiceOver describing surroundings and onscreen content in richer detail, but we did not read a processing statement for those.

10. Context suggestions: Magic Cue, Now Nudge and Call Context

All three vendors describe features that read your context and offer a suggestion. Google's Pixel 10 post says Magic Cue 'connects information across apps like Gmail, Calendar, and Messages' and 'runs securely and privately through a combination of the Tensor G5 chip and Gemini Nano.' Samsung says Now Nudge 'recognizes what's on your screen and offers context-based suggestions,' such as opening Calendar while texting about plans. Apple says Call Context can 'proactively surface relevant information from across apps' when you call a business.

Limit to watch: these features work because they look across your apps. That is the point, and also the privacy trade-off. Only Google's page, of the three we read, says anything about where it runs.

What privacy questions should I ask before switching one on?

For longer-lived data, such as chat history, see our guide to questions about AI data retention. To learn how to judge a vendor's wording, read how to choose an AI tool.

What are the limitations of this list?

We read vendor pages, not devices. We did not test whether any feature works as described, whether a privacy statement is accurate, or how a feature behaves on your phone and software version. Pages differ by region and version and change without notice. We chose features that appear on the vendors' own pages and are widely discussed; this is a selection, not a ranking. If a statement here is out of date, tell us through the corrections page and see our methodology for how we label evidence.